01
Permission is part of architecture
An agent that can propose a change and an agent that can execute a destructive command are different systems. The workflow should make that boundary explicit.
C2
Public proof / open source
Building guardrails for AI-assisted engineering
AgentLint is a public, MIT-licensed project created to catch quality, security, and infrastructure risks while coding agents work—not after the session has drifted.
77
public rules
8
rule packs
MIT
license
2.5.5
audited public release
The operating problem
Long agent sessions create invisible drift.
AI coding tools can increase local output while simultaneously increasing the number of changes, trust boundaries, and review decisions a team must manage.
AgentLint turns recurring failure modes into executable checks: secrets, destructive operations, missing tests, infrastructure risk, session quality, and other agent-specific behavior.
The project is evidence of implementation and systems thinking. It is not evidence for a universal productivity multiplier, and this site deliberately makes no such claim.
Transferable lessons
02
Guardrails belong near the action
A policy document is too far from a secret, skipped test, dangerous infrastructure edit, or destructive command. Controls should run where work happens.
03
Measure before claiming acceleration
Adoption is not evidence of productivity. Establish a baseline, define the workflow and cohort, and inspect quality and rework alongside speed.
Public proof / ConstraintLoop
Completion should be an evidence decision.
ConstraintLoop is an MIT-licensed framework for explicit, versioned completion contracts made from tests, static checks, metrics, artifacts, and optional independent model rubrics.
It supports Claude Code, Codex, and Gemini CLI while keeping CI authoritative. Its convergence loops are bounded by repair, unchanged-result, and time budgets rather than an open-ended instruction to keep trying.